Compaction is when Claude Code summarizes a long conversation so that it fits in the model's working space. One way of doing this reuses the prompt cache, a stored copy of the conversation that the API can pick up again cheaply.
That path now reports progress while the summary streams in. The indicator shows how much of the response has arrived and that the model is responding. The fallback summarizer now also shows a "requesting" state at the start of every attempt, not only on retries.
The compaction indicator should stop looking frozen while the cache-reusing summary is being written.
tengu_compact_cache_prefix Off in both readingsThe flag server returned off for the account this site reads and for the anonymous baseline. A reading of off cannot rule out a rollout these two readings sit outside of.
This account: off · anonymous baseline: off · compiled default in v2.1.283: on
These values were read against a different version of Claude Code, so treat them as the nearest reading available instead of one taken on this release.
Read once, for one account on one subscription tier, against v2.1.283. It isn't a statement about your account. What a flag value here can and cannot tell you